JP5957269B2 - Voice recognition server integration apparatus and voice recognition server integration method - Google Patents

Description

  The present invention connects a terminal device for a user to operate using voice and a plurality of voice recognition servers that recognize voice data and return the result, and integrates the recognition results obtained by the plurality of voice recognition servers. The present invention relates to an apparatus and a method for providing a user with an optimal speech recognition result.

  A number of devices are equipped with a voice recognition function for performing operations on in-vehicle information devices and mobile phones by the voice of the user. Furthermore, in recent years, with the development of data communication technology, a method of transmitting voice data to a server and performing more accurate voice recognition using abundant calculation resources of the server is becoming widespread. Also, as shown in Patent Document 1, in order to use such a voice recognition server more effectively, a method is proposed in which a relay server is placed between the personal terminal and the voice recognition server, and additional processing is performed therein. Has been.

JP 2008-242067 A Japanese Patent Laid-Open No. 2002-116796 JP 2010-224301 A

  An example in which a speech recognition server is operated as a general-purpose service is increasing, and there are many cases where the entity that provides services for terminals handled by users and the entity that operates the speech recognition server are different. Even when the operating entities are the same, the development of the speech recognition server and the development of the speech recognition application are performed separately, and they may not be mutually optimized. In such a situation, a general-purpose speech recognition server may generally not exhibit sufficient performance for a specific phrase while exhibiting high performance.

  On the other hand, when focusing on a specific user who uses a specific application, there are words that are not general but highly important, such as names of acquaintances of the user and names of favorite music. In order to recognize such words and phrases, it is desirable to provide a dedicated speech recognition server. However, it is often impossible to develop a dedicated speech recognition server at a sufficient cost. It will be inferior to the recognition server. As described above, the general-purpose speech recognition server and the dedicated speech recognition server are not good at recognizing depending on phrases and have different speech recognition performance. Therefore, it is required to use these properly depending on the words uttered by the user, but since speech recognition is the task of "estimating the content for utterances whose contents are unknown", it is necessary to know the utterance content in advance and use the server properly It is impossible in principle.

  An object of the present invention is to optimally integrate the speech recognition result of the general-purpose speech recognition server and the speech recognition result of the dedicated speech recognition server, and finally provide a speech recognition function with few errors.

  In the present invention, a list of specific words included in a terminal device possessed by a user is obtained in advance, and a dedicated speech recognition server is constructed based on the data of those words. Moreover, the performance of a general-purpose speech recognition server is evaluated in advance using those data. Based on the evaluation results, which of the recognition results obtained from the dedicated and general-purpose speech recognition servers is used and what weighting is applied to them, the optimal recognition result can be obtained. Keep it in the form of a database. When the user actually uses the speech recognition function, the input speech is recognized by a dedicated and general-purpose speech recognition server, and then the results obtained there are compared with the contents of the database described above to obtain the optimal speech. Get recognition result. Further, by using the response speed in addition to the correctness of the speech recognition as a reference for the prior evaluation, it is possible to obtain the correct speech recognition result as soon as possible.

If an example is given to the speech recognition server integration device of the present invention, a device that relays between a terminal device for a user to operate using speech and a speech recognition server that recognizes speech data and returns the result. a is, an integrated system learning unit for the user to save learns recognition result integration parameter based on a list of phrases frequently used words and phrases or user has registered, speech data generated by a user with the intention of speech recognition means for transmitting the audio data and means for receiving from the terminal device, the audio data thus received to the general-purpose speech recognition server and the specialized speech recognition server is, by the general-purpose speech recognition server and the specialized speech recognition server of the audio data means for receiving the recognition result, the recognition result by a general purpose speech recognition server and the specialized speech recognition server, the stored recognition result integration Compared with parameters, those comprising a recognition result integration unit that selects an optimal recognition result, and means for transmitting the selected recognition result to the terminal device.

  In the speech recognition server integration device of the present invention, further, means for receiving a word registered by the user or a list of words frequently used by the user from the terminal device, and speech synthesis for generating synthesized speech based on the received word Unit, means for transmitting the generated synthesized speech to the general-purpose speech recognition server and the dedicated speech recognition server, and means for receiving a recognition result of the synthesized speech by the general-purpose speech recognition server and the dedicated speech recognition server, The integration method learning unit may analyze the phrase that is the basis of the synthesized speech and the recognition result together to learn and store the recognition result integration parameter.

In the speech recognition server integration device of the present invention, means for receiving a list of phrases registered by the user or frequently used phrases from the terminal device, and a list of phrases for recognition from the general-purpose speech recognition server And a phrase comparison / similarity estimation unit that compares the phrase list for recognition with the phrase list received from the terminal device and estimates similarity, and the integration method learning unit it may be those stores as a recognition result integration parameter.

If an example of a speech recognition server integration process of the present invention, the steps of the user and stores the learned recognition result integration parameter based on a list of phrases frequently used words and phrases or user has registered, the user speech recognition Transmitting voice data that is voice data intended to be transmitted to a general-purpose voice recognition server and a dedicated voice recognition server, and receiving a recognition result of the voice data by the general-purpose voice recognition server and the dedicated voice recognition server And comparing the recognition result of the general-purpose speech recognition server and the recognition result of the dedicated speech recognition server with the recognition result integration parameter and selecting the optimum speech recognition result.

  According to the present invention, the recognition result of the general-purpose speech recognition server is regarded as important for general words and phrases, and the result of the dedicated speech recognition server is regarded as important for user-specific words and phrases in an optimum form for each input. The recognition results are integrated, and finally it is possible to provide the user with a voice recognition function with few errors. In addition to fewer errors, it is possible to realize a highly convenient system in terms of response speed.

It is a block diagram of the speech recognition server integration apparatus of Example 1 of this invention. It is a figure which shows the process of parameter estimation for result integration using the speech synthesis of Example 1 of this invention. It is a figure which shows an example of the parameter for result integration using the single general purpose speech recognition server of this invention. It is a figure which shows an example of the parameter for result integration using the multiple general purpose speech recognition server of this invention. It is a figure which shows an example of the integration method of the recognition result of the some server of Example 1 of this invention. It is a figure which shows an example of the parameter for result integration using the recognition result reliability of the multiple general purpose speech recognition server of this invention. It is a figure which shows an example of the parameter for result integration using the recognition result reliability and misrecognition result of the multiple general purpose speech recognition server of this invention. It is a figure which shows an example of the integration method of the recognition result using the conversion of the same phonetic notation of this invention. It is a figure which shows the structural example of the user terminal for implement | achieving this invention. It is a figure which shows an example of the creation method of the user dictionary in this invention. It is a figure which shows an example of a structure of the speech synthesis part in this invention. It is a figure which shows an example of the parameter for result integration which considered the response time of this invention. It is a block diagram of the speech recognition server integration apparatus of Example 2 of this invention. It is a block diagram of the speech recognition server integration apparatus of Example 3 of this invention. It is a figure which shows the process of parameter estimation for result integration using the recognition term phrase list of Example 3 of this invention. It is a block diagram of the speech recognition server apparatus of Example 4 of this invention. It is a block diagram of the speech recognition server apparatus of Example 5 of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. Note that components having the same function are denoted by the same names and reference symbols throughout the drawings for describing the embodiments for carrying out the invention, and the repetitive description thereof will be omitted.

  FIG. 1 is a diagram illustrating a configuration example of a speech recognition server integration device based on Embodiment 1 of the present invention. The voice recognition function is provided using the user terminal 102, the relay server 104, the general-purpose voice recognition server group 106, and the dedicated voice recognition server 108. The general-purpose speech recognition server group 106 may be a single general-purpose speech recognition server.

  The user terminal 102 is a terminal device possessed by the individual user. In addition to obtaining input voice data and providing a service based on the voice recognition result, the user terminal 102 holds a phrase list unique to the user such as an address book and a song name list. Yes. Hereinafter, the phrase list unique to these users is referred to as a “user dictionary”. The user dictionary holds a list of words registered by the user or frequently used by the user.

  The general-purpose speech recognition server group 106 is one or more speech recognition servers that are not assumed to be used only by services realized by the present invention. In general, a large-scale word list is built in and recognition performance for various words is high, while some words included in the user dictionary may not be recognized correctly.

  The dedicated speech recognition server 108 is a speech recognition server specialized for services realized by the present invention, and is designed to recognize all or most of the words included in the user dictionary. The dedicated speech recognition server 108 is designed to output a result of “no recognition result” when a phrase not included in the user dictionary is input. The dedicated speech recognition server is not limited to being configured as a server, and may be a dedicated speech recognition device, or may be built in a user terminal or a relay server as in the second and fifth embodiments.

  The relay server 104 corresponds to the “voice recognition server integration device” of the present invention, and connects the user terminal 102 and the voice recognition servers 106 and 108 to integrate voice recognition results. Data exchange with the user terminal 102 is performed via the terminal device communication unit 110. Data exchange with the voice recognition servers 106 and 108 is performed via the recognition server communication unit 112. The relay server 104 includes a terminal device communication unit 110, a speech synthesis unit 114, an integration method learning unit 116, a signal processing unit 120, a recognition result integration unit 122, a recognition server communication unit 112, and the like.

  The operation of the relay server 104 will be described. First, when the user sets the user terminal 102 in a communicable state, data in the user dictionary 124 is transmitted via the terminal device communication unit 110. This data is sent directly to the recognition server communication unit 112 and further sent to the dedicated speech recognition server 108. The dedicated speech recognition server 108 performs tuning so as to correctly recognize the words included in the user dictionary data sent thereto. On the other hand, the user dictionary data received by the terminal device communication unit 110 is also sent to the speech synthesis unit 114. Here, synthesized speech data is created based on user dictionary data sent as a character string. The synthesized speech data for one word may be one or a plurality of pieces having different sound quality. The generated synthesized speech data is sent to the general-purpose speech recognition server group 106 and the dedicated speech recognition server 108 via the recognition server communication unit 112. When the recognition results for these are returned from each server, the recognition server communication unit 112 receives them and sends them to the integrated method learning unit 116. The integration method learning unit 116 analyzes the user dictionary data that is the basis of the synthesized speech and the recognition result, and learns parameters for integrating the recognition result. The obtained parameters are stored as result integration parameters 118. At this point, the pre-learning process of the system using the present invention ends.

  When the user actually uses the voice interface, the input voice data acquired by the user terminal 102 is received by the terminal device communication unit 110. The received data is sent to the signal processing unit 120 and subjected to necessary processing. Here, the necessary processing refers to, for example, removing noise from input speech including noise, but is not necessarily required, and no processing may be performed. Data output from the signal processing unit 120 is sent to the general-purpose speech recognition server group 106 and the dedicated speech recognition server 108 via the recognition server communication unit 112. The recognition results returned from these servers are sent to the recognition result integration unit 122 via the recognition server communication unit 112. The recognition result integration unit 122 compares a plurality of recognition results with the parameters included in the result integration parameter 118, and selects an optimal recognition result. The selected recognition result is sent to the user terminal 102 via the terminal device communication unit 110. Based on this result, the user terminal 102 provides services such as setting a destination of a navigation function, making a call, and playing a music piece.

  FIG. 2 is a diagram showing a processing procedure until the result integration parameter is created using the user dictionary data in the configuration shown in FIG. First, user dictionary data is sent to a dedicated speech recognition server as it is. In the dedicated speech recognition server, the speech recognition engine is tuned so that the sent phrase is the recognition target. Therefore, when data uttering a word that is not included in the user dictionary is sent, the dedicated speech recognition server returns an incorrect result or returns a result indicating that recognition is impossible. On the other hand, the user dictionary data is also sent to the speech synthesis unit, where synthesized speech data is created. Normally, one synthesized speech is created for one word, but if the speech synthesizer has a function that allows you to select the speaker, speech speed, voice pitch, etc., change them. Thus, if a plurality of synthesized speech data is created for the same word / phrase, it is possible to further improve the performance of the integrated method learning performed later.

  The synthesized speech data obtained in this way is sent to each general-purpose speech recognition server and dedicated speech recognition server. Recognition results are returned from those servers. In addition, not only the recognition result but also a reliability score associated therewith may be attributed together. Based on these, the integration method learning unit learns the integration method, and the result is stored in a result integration parameter.

  FIG. 3 is a diagram illustrating an example of the simplest configuration of result integration parameters. In this example, it is assumed that there is only one general-purpose speech recognition server, and the server holds only whether or not each word / phrase in the user dictionary is correctly recognized by ○ and ×. That is, this figure shows that the words “Ichiro Suzuki” and “Jiro Yamada” were correctly recognized by the general-purpose speech recognition server but were not correctly recognized otherwise. FIG. 4 is an example when similar learning is performed using three general-purpose speech recognition servers.

  FIG. 5 shows a processing procedure when the recognition is actually performed using the results shown in FIGS. Input voice data is first pre-processed by the signal processing unit. As a typical process in the signal processing unit, there is a noise suppression process as disclosed in Patent Document 1. As a result of processing in the signal processing unit, one piece of sound data is usually obtained for one piece of input sound data, but there are cases where a plurality of pieces of sound data are obtained by changing the setting. In such a case, the processing described below is repeated as many times as the number of audio data. Further, when it is considered that the processing in the signal processing unit is unnecessary, the input voice data is directly used as output data of the signal processing unit.

  The output data of the signal processing unit is sent to the general-purpose speech recognition server and the dedicated speech recognition server. All these results are sent to the recognition result integration unit. The recognition result integration unit first checks the recognition result of the dedicated speech recognition server. If the recognition result of the dedicated recognition server is “no recognition result”, the final recognition result is determined only from the recognition result of the general-purpose speech recognition server. That is, when there is only one general-purpose speech recognition server, the result is adopted as it is. If there are multiple units, a majority vote is taken between the recognition results. When taking a majority vote, if each recognition server gives a reliability score, a majority vote weighted by that value can be used. In addition, the performance of each recognition server can be estimated in advance and used as a weighting coefficient. For such integration of recognition results of a plurality of speech recognition servers for general words / phrases, a known technique as shown in Patent Document 2 can be used.

  On the other hand, when a phrase included in the user dictionary data is obtained as a recognition result of the dedicated speech recognition server, the result integration parameter as shown in FIG. 3 or FIG. 4 is referred to. For example, in the example of FIG. 3, when the recognition result of the dedicated speech recognition server is “Hitachi Taro”, when the corresponding line of the result integration parameter is viewed, this phrase should not be recognized by the general-purpose speech recognition server. Therefore, the result of the dedicated speech recognition server is adopted as it is. On the other hand, when the recognition result of the dedicated speech recognition server is “Ichiro Suzuki”, it can be seen from the corresponding line of the result integration parameter that this phrase can also be recognized by the dedicated speech recognition server. Then, the recognition result of the general-purpose speech recognition server is checked next. If the recognition result of the general-purpose speech recognition server is also “Ichiro Suzuki”, “Ichiro Suzuki” can be used as the final recognition result, but if not, the general-purpose performance generally seems to be high. The result of the voice recognition server is given priority, or the recognition result of the general-purpose voice recognition server and the dedicated voice recognition server having the higher reliability score is adopted as the final recognition result. As a result, even if a word of pronunciation similar to “Ichiro Suzuki” is misrecognized by the dedicated speech recognition server, it can be rejected based on the recognition result of the general-purpose speech recognition server. Become. The same applies to the example of FIG. 4, and for “Hitachi Taro”, the result of the dedicated speech recognition server is used unconditionally. Since “Ichiro Suzuki” is a word that can be recognized by all three general-purpose speech recognition servers, the final recognition result is determined by a majority decision based on these recognition results, or a majority decision including a dedicated speech recognition server. To do. If the recognition result of the dedicated speech recognition server is “Jiro Yamada”, the only general-purpose speech recognition server that can correctly recognize this is the first, so this server and the dedicated speech recognition server The final recognition result is obtained by performing the same processing as in the example of FIG.

  FIG. 6 shows another implementation example of result integration parameters different from those in FIGS. 3 and 4. Here, when a certain word / phrase is recognizable by each general-purpose speech recognition server, the probability that the word / phrase is correctly recognized is replaced with a numerical value of weight and held. Here, the probability of correct recognition is, for example, that the word “Ichiro Suzuki” is recognized with synthesized speech made by changing various parameters for speech synthesis, and the number of recognition results is correct. It can be estimated depending on whether there was. In addition, when the general-purpose speech recognition server is configured to return a plurality of recognition result candidates, the average rank of the correct words, the average reliability score, and the like can be used. A value obtained by converting these values into weight values by appropriate non-linear conversion is held in a result integration parameter. In this example, when the recognition result of the dedicated speech recognition server is “Ichiro Suzuki”, the result of the general-purpose server 1 is “Ichiro Sasaki”, and the results of the general-purpose servers 2 and 3 are “Ichiro Suzuki”, the weight of “Ichiro Sasaki” Is 3.0 and the weight of “Ichiro Suzuki” is the sum of 1.4 and 1.2, which is 2.6, and the former is larger, so “Ichiro Sasaki” is the final recognition result.

  FIG. 7 shows another implementation example of result integration parameters different from those shown in FIGS. Here, even when a word / phrase included in the user dictionary data is recognized by the general-purpose speech recognition server and not correctly recognized, the recognition result obtained at that time is stored as a result integration parameter. Setting the weight of each server is the same as in the example of FIG. When a plurality of experiments are performed, only the most frequent result or a plurality of recognition results may be stored. In addition, the recognition result of the second place or less may be stored together regardless of the number of experiments. When executing the recognition, the result integration parameter is referred to based on the recognition result of the dedicated speech recognition server, as in the previous examples. At this time, it is checked whether or not the recognition result of the general-purpose speech recognition server matches that stored in the result integration parameter. For example, the recognition result of the dedicated speech recognition server is “Hitachi Taro”, the result of the general-purpose server 1 is “Hitachi City”, the result of the general-purpose server 2 is “20 years old”, and the result of the general-purpose server 3 is “Hitachi”. In this case, after converting the result of the general-purpose server 1 into “Hitachi Taro”, a majority decision is made on each recognition result, and finally “Hitachi Taro” is selected.

  FIG. 8 is a diagram illustrating an example of a speech recognition result integration method using detection of homonymous notation. As shown in the figure, when the recognition result of the dedicated speech recognition server is “Ichiro Sato”, it is compared with each recognition result of the general-purpose speech recognition server to check whether or not the same phonetic notation is included. Here, in order to estimate the pronunciation from the notation, in the case of Japanese, each kanji reading is held as data, and the pronunciation notation is obtained by concatenating the kanji readings constituting the word. In the case of English, the rules for giving readings for partial spellings are retained, and phonetic notation is obtained by sequentially applying these rules. Even in the case of other languages, it is well known that pronunciation notation can be obtained by a technique generally called Grapheme to Phoneme. In some cases, pronunciation information is included in the user dictionary data such as kanji notation and kana notation, and in such a case, it is utilized. If the above-mentioned check includes homonym notation, the recognition result notation is converted into a recognition result notation by the dedicated speech recognition server. In the example in the figure, since the recognition result “Ichiro Sato” of the general-purpose speech recognition server 1 is the same sound as the recognition result of the dedicated speech recognition server, it is converted to “Ichiro Sato”. As a result, the result of the majority decision by the three general-purpose speech recognition servers is “Ichiro Sato”, which is adopted as the final result.

  FIG. 9 is a diagram showing an example of a specific implementation form of the user terminal, taking as an example the case of providing a navigation function, a hands-free call function, etc. in an automobile. In (a), all functions such as a microphone device 904, an application 906, and a communication module 908 are mounted in the car navigation device 902. In (b), the car navigation device 902 and the smartphone 910 are connected, the microphone device 904 uses the car navigation device 902, and the communication unit 908 uses the smartphone 910. The applications 912 and 914 are arranged in a distributed manner in the car navigation apparatus and in the smartphone, or arranged in only one of them according to the respective functions. In (c), all functions are implemented in the smartphone 910.

  FIG. 10 is a diagram showing an example of a method for creating the user dictionary 124 constituting the present invention. For example, when the address book 1002 exists in the user terminal 102, the names of people included therein are registered in the user dictionary. Similarly, when there is a music list 1004 of the music player, the music name and artist name included therein are registered in the user dictionary. Also, the page title registered as the bookmark 1006 of the web browser can be registered in the user dictionary. In addition, a method of analyzing data such as mail 1008 and short messages stored in the user terminal and registering frequently occurring phrases in the user dictionary is also possible. With respect to these data, when the user terminal is connected to the system according to the present invention for the first time, in addition to transmitting all user dictionary data contained in the user terminal to the system, new to the address book, music list, etc. At the time of adding an entry, it is also possible to adopt a method in which only newly added data is additionally transmitted to the system and the result integration parameter is prompted to be updated. At this time, it is necessary to update not only the result integration parameter but also the collation dictionary of the dedicated speech recognition unit at the same time.

  FIG. 11 is a diagram showing an example of a special configuration specialized in the present invention by changing the configuration of a general speech synthesis unit. Generally, the speech synthesizer 114 includes a synthesized speech creation unit 1102 and speech unit data 1106 to 1110. Here, segment data is the name of data for use in a method for creating synthesized speech by directly joining data, but instead of directly joining, a method for synthesizing waveforms by statistical processing and signal processing is used. Even in this case, since a similar data set is used for processing units such as individual phonemes and syllables, the following method can be applied. A synthesized speech creating unit 1102 joins speech unit data, performs appropriate signal processing if necessary, and creates standard synthesized speech. However, in the present invention, since it is important to know how each general-purpose speech recognition server group reacts to the voice of the specific user who is the owner of the user terminal, it is created by the speech synthesizer. The synthesized voice is also preferably similar to the voice of the user. Therefore, every time the user uses the voice recognition function or every time other voice function or voice call is used, the voice is stored as user voice data 1112, and this is used to make a voice conversion unit 1104. Thus, the standard synthesized speech is converted to user adaptive speech. By using the converted speech as an input to the general-purpose speech recognition server group, it is possible to perform performance prediction with higher accuracy, and it can be expected that the value of the result integration parameter will be more appropriate.

  FIG. 12 is a diagram illustrating an example of result integration parameters when the response speed is used as an evaluation criterion in addition to correct speech recognition. In this example, recognition using synthesized speech corresponding to each word / phrase included in the user dictionary data is executed, and the average time required for the processing is stored as a parameter. In this example, if the recognition result of the dedicated speech recognition server is “Ichiro Suzuki”, the recognition result of the general-purpose server 2 is expected to be obtained in 0.5 seconds. You have to wait 1.5 seconds to get it. When this response time exceeds the upper limit assumed by the application, the result integration process is performed when the result of the general-purpose server 2 is obtained. As a result, assuming that the result integration process takes almost no time, the final recognition result can be obtained with a response time of about 0.5 seconds, and the convenience for the user can be improved.

  FIG. 13 is a diagram showing a configuration of an example in which a function equivalent to the example shown in FIG. 1 is realized using the dedicated speech recognition unit 108 incorporated in the user terminal. Here, the user terminal 102 recognizes a word / phrase included in the user dictionary 124 using the dedicated voice recognition unit 108 existing inside without using the relay server 104. The method for performing the performance evaluation of the general-purpose speech recognition server group 106 in advance using the user dictionary data is the same as that shown in FIG. At the time of recognition execution, recognition by the general-purpose speech recognition server 106 is performed via the relay server 104, and at the same time, recognition is also performed by the dedicated speech recognition unit 108 in the user terminal. Such a method of using a voice recognition unit in a terminal and a voice recognition unit connected via a communication device is also shown in Patent Document 3, but the invention described in Patent Document 3 While the results are selected based on whether the route is established or not, the present invention is different in that the result integration parameter obtained based on the result of speech recognition performed in advance is used. Yes.

  FIG. 14 is a diagram showing another configuration example of the speech recognition server integration device according to the present invention. Here, as a function of the general-purpose speech recognition server group 106, a case is assumed in which a recognition term phrase list used therein is available. Under such conditions, the user dictionary data sent from the user terminal 102 to the relay server 104 is sent to the phrase comparison / similarity estimation unit 126. This section compares the recognition phrase list obtained from the general-purpose speech recognition server group 106 with the user dictionary data, and determines whether each word included in the user dictionary 124 can be correctly recognized by each server. The determination result is sent to the integration method learning unit 116, and the result organized as a parameter is held in the result integration parameter 118. On the other hand, the user dictionary data is sent to the dedicated speech recognition server 108 as it is, and the dedicated speech recognition server is tuned as in the example shown in FIG.

  When input voice data is sent from the user terminal 102 in such a prepared state, the general-purpose voice recognition server 106 and the dedicated voice data are sent via the signal processing unit 120 as in the example shown in FIG. The data is sent to the voice recognition server 108. The recognition results returned from these servers are sent to the recognition result integration unit 122, where an optimum recognition result is selected by comparison with the result integration parameter 118. After the selected recognition result is transmitted to the user terminal 102, it is the same as the example shown in FIG.

  FIG. 15 is a diagram showing a processing procedure until the result integration parameter is created using the user dictionary data in the configuration shown in FIG. In this example, a recognition speech phrase list is simply acquired from each general-purpose speech recognition server without creating synthesized speech or trying to execute speech recognition using it. These lists are compared with the words / phrases included in the user dictionary data, and data indicating which word / phrase list of which general-purpose speech recognition server contains each word / phrase in the user dictionary data. Here, since there can only be either (O) or not (X) included in the recognition term phrase list, the result integration parameters that summarize the obtained results are shown in FIG. It becomes the same as FIG. Therefore, how to use the actual recognition is the same as the example described above. Further, when it is possible to obtain not only the phrase list but also a language model representing the ease of recognition of these phrases from each general-purpose speech recognition server, the weighted result integration parameters as shown in FIG. Can also be created. For example, when an N-gram language model that is a typical language model is used, the unigram value is set to be easy to recognize the word, or the double gram or trigram maximum value is set to be easy to recognize the word. A method is conceivable.

  FIG. 16 shows an example configuration in which a function equivalent to the example shown in FIG. 1 is realized by a device in which an input / output function with a user and a voice recognition server integration function are incorporated in a single device. FIG. Here, the user dictionary data included in the user dictionary 124 stored inside the speech recognition server integration device 104 is transferred to the speech synthesis unit 114 and the recognition server communication unit 112 in the device. The voice spoken by the user is captured by the microphone device 128 and transferred to the signal processing unit 120. The processing method using these is the same as that described in the example of FIG. 1, and as a result, the recognition result integration unit 122 finalizes the recognition result. This recognition result is transferred to the display unit 132 in the apparatus and presented to the user.

  FIG. 17 is a diagram showing a configuration when the functions of the dedicated speech recognition server are incorporated in the speech recognition server integration device based on the example shown in FIG. The part in which the input voice is taken in from the microphone device 128 included in the voice recognition server integration apparatus 104 and the user dictionary data is transferred from the user dictionary 124 is the same as the example of FIG. The unit 108 is incorporated in the apparatus, and directly reads out the contents of the user dictionary and recognizes audio data sent from the microphone apparatus. The single recognition result obtained there is sent to the recognition result integration unit 122 and integrated with the recognition result obtained by the general-purpose speech recognition server group 106. The integrated recognition result is sent to the application 130 existing in the apparatus, where it is utilized according to the purpose of each application.

  INDUSTRIAL APPLICABILITY The present invention can be used as a voice data relay device that is interposed between an in-vehicle terminal and a voice recognition server to provide a highly accurate voice recognition function.

102 User terminal 104 Relay server 106 General-purpose speech recognition server group 108 Dedicated speech recognition server 110 Terminal device communication unit 112 Recognition server communication unit 114 Speech synthesis unit 116 Integration method learning unit 118 Result integration parameter 120 Signal processing unit 122 Recognition result integration unit 124 User dictionary 126 Phrase comparison / similarity estimation unit 128 Microphone device 130 Application 132 Display unit

Claims (15)

A device that relays between a terminal device for a user to operate using voice and a voice recognition server that recognizes voice data and returns the result,
And integration method learning unit for the user to save learns recognition result integration parameter based on a list of phrases frequently used words and phrases or user registered,
Means for receiving, from the terminal device, voice data that is voice data intended by a user for voice recognition;
Means for transmitting the received voice data to a general-purpose voice recognition server and a dedicated voice recognition server;
Means for receiving a recognition result of the voice data by the general-purpose voice recognition server and a dedicated voice recognition server;
The recognition result of the general-purpose speech recognition server and the specialized speech recognition server, as compared with the stored recognition result integration parameters, and the recognition result integration unit that selects an optimal recognition result,
A voice recognition server integration device comprising: means for transmitting the selected recognition result to the terminal device. The speech recognition server integration device according to claim 1, further comprising:
Means for receiving a list of phrases registered by the user or frequently used by the user from the terminal device;
A speech synthesizer that generates synthesized speech based on the received phrases;
Means for transmitting the generated synthesized speech to the general-purpose speech recognition server and the dedicated speech recognition server;
Means for receiving recognition results of the synthesized speech by the general-purpose speech recognition server and the dedicated speech recognition server;
The speech recognition server integration device, wherein the integration method learning unit analyzes a phrase that is a basis of the synthesized speech and the recognition result, learns and stores a recognition result integration parameter. The speech recognition server integration device according to claim 1, further comprising:
Means for receiving a list of phrases registered by the user or frequently used by the user from the terminal device;
Means for receiving a word list for recognition from the general-purpose speech recognition server;
A phrase comparison / similarity estimation unit that compares the phrase list for recognition with the phrase list received from the terminal device and estimates similarity; and
The integration method learning section, the estimation result speech recognition server integration device, characterized in that the stored as a recognition result integration parameter. A device for a user to perform operations using voice,
And integration method learning unit for the user to save learns recognition result integration parameter based on a list of phrases frequently used words and phrases or user registered,
Means for transmitting voice data , which is voice data intended by the user for voice recognition, to the general-purpose voice recognition server and the dedicated voice recognition server;
Means for receiving a recognition result of the voice data by the general-purpose voice recognition server and a dedicated voice recognition server;
A recognition result integration unit that compares the recognition results of the general-purpose speech recognition server and the dedicated speech recognition server with the stored recognition result integration parameter and selects an optimal recognition result;
A voice recognition server integration device comprising a display unit for displaying the selected recognition result. The voice recognition server integration device according to claim 4 , further comprising:
A user dictionary for storing words registered by the user or frequently used by the user;
A speech synthesizer that generates synthesized speech based on words stored in the user dictionary;
Means for transmitting the generated synthesized speech to the general-purpose speech recognition server and the dedicated speech recognition server;
Means for receiving recognition results of the synthesized speech by the general-purpose speech recognition server and the dedicated speech recognition server;
The speech recognition server integration device, wherein the integration method learning unit analyzes a phrase that is a basis of the synthesized speech and the recognition result, learns and stores a recognition result integration parameter. The voice recognition server integration device according to claim 4 , further comprising:
A user dictionary that stores a list of words registered by the user or frequently used by the user;
Means for receiving a word list for recognition from the general-purpose speech recognition server;
A phrase comparison / similarity estimation unit that compares the phrase list for recognition with the phrase list of the user dictionary and estimates similarity ;
The integration method learning section, the estimation result speech recognition server integration device, characterized in that the stored as a recognition result integration parameter. In the speech recognition server integration device according to any one of claims 1 to 6,
The dedicated speech recognition server creates a recognition target word list based on a word registered by the user or a list of words frequently used by the user, and can recognize the words included in the list with high accuracy. A featured voice recognition server integration device. In the speech recognition server integration device according to any one of claims 1 to 6,
The dedicated speech recognition server, wherein the dedicated speech recognition server is incorporated in the speech recognition server integration device or terminal device as a dedicated speech recognition unit. In the speech recognition server integration device according to claim 2 or 5,
The recognition result integration parameter stores the correctness of the recognition result of the general-purpose speech recognition server for a word registered by the user or a word frequently used by the user,
The recognition result integration unit takes out the correctness of the recognition result of the general-purpose speech recognition server with respect to the recognition result by the dedicated speech recognition server from the recognition result integration parameter based on the recognition result by the dedicated speech recognition server , speech recognition server integration device and selects an optimum recognition result based on the speech recognition result serial result fetch by a general purpose speech recognition server such that positive. In the speech recognition server integration device according to claim 2 or 5,
The recognition result integration parameter is a value representing whether the recognition result of the general-purpose speech recognition server for the words registered by the user or frequently used by the user is correct, and the reliability of the recognition result of the general-purpose speech recognition server for each word Is to save
The recognition result integration unit, based on the recognition result by the dedicated speech recognition server , correctness of the recognition result of the general-purpose speech recognition server with respect to the recognition result by the dedicated speech recognition server from the recognition result integration parameter and its reliability It was removed, and in that the recognition result extraction to select a generic speech recognition by the server Ruoto voice recognition result to the reliability weighting results optimal recognition based on a result of integration by the like is positive A featured voice recognition server integration device. In the speech recognition server integration device according to claim 2 or 5,
The recognition result integration parameter is for measuring the time taken for recognition of a dedicated and general-purpose speech recognition server for a word registered by the user or a word frequently used by the user, and storing a measurement value
The recognition result integration unit, based on the recognition result by the dedicated speech recognition server , extracts the recognition time of the dedicated and general-purpose speech recognition server for the recognition result by the dedicated speech recognition server from the recognition result integration parameter, Out of the dedicated and general-purpose speech recognition servers, the recognition result is extracted based on the recognition result that is less than the upper limit of the recognition time determined depending on the application, so that the optimum recognition can be performed based on the extracted recognition result. A speech recognition server integration device, wherein a result is selected. In the speech recognition server integration device according to claim 2 or 5,
The recognition result integration parameter stores a recognition result including a misrecognition result of a general-purpose speech recognition server for a word registered by a user or a word frequently used by a user,
The recognition result integration unit, based on the recognition result by the dedicated speech recognition server, obtains a recognition result including an erroneous recognition result of the general-purpose speech recognition server for a recognition result by the dedicated speech recognition server from the recognition result integration parameter. If the speech recognition result of the general-purpose speech recognition server matches the recognition result including the extracted misrecognition result, it is converted into a correct word and the most suitable recognition result is selected by performing a majority decision on each recognition result. A speech recognition server integration device characterized by the above. A step for the user to store learning the recognition result integration parameter based on a list of phrases frequently used words and phrases or user registered,
Transmitting voice data , which is voice data intended by the user for voice recognition, to the general-purpose voice recognition server and the dedicated voice recognition server;
Receiving a recognition result of the voice data by the general-purpose voice recognition server and a dedicated voice recognition server;
Comparing the recognition result of the general-purpose speech recognition server and the recognition result of the dedicated speech recognition server with the recognition result integration parameter, and selecting the optimum speech recognition result;
A speech recognition server integration method comprising: 14. The voice recognition server integration method according to claim 13, further comprising:
Generating synthesized speech based on words registered by the user or frequently used by the user;
Transmitting the generated synthesized speech to the general-purpose speech recognition server and a dedicated speech recognition server;
Receiving a recognition result of the synthesized speech by the general-purpose speech recognition server and a dedicated speech recognition server,
Step of storing learning the recognition result integration parameter, the synthesized speech of the combined group and became the phrase the recognition result and analyzed, characterized in that the store to learn recognition result integration parameter Speech recognition server integration method. 14. The voice recognition server integration method according to claim 13, further comprising:
Obtaining a list of phrases registered by the user or frequently used by the user;
Receiving a recognition word list from the general-purpose speech recognition server;
Comparing the recognition word list with a word registered by the user or a list of words frequently used by the user, and estimating a similarity,
Step, the speech recognition server integration method characterized by storing the estimation result as a recognition result integration parameter to be saved by learning the recognition result integration parameter.

Images